In [3]:
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)

# Input data files are available in the read-only "../input/" directory
# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory

import os
for dirname, _, filenames in os.walk('/content'):
    for filename in filenames:
        print(os.path.join(dirname, filename))

/content/.config/.last_opt_in_prompt.yaml
/content/.config/gce
/content/.config/.last_survey_prompt.yaml
/content/.config/.last_update_check.json
/content/.config/config_sentinel
/content/.config/default_configs.db
/content/.config/active_config
/content/.config/logs/2024.06.14/17.38.49.910604.log
/content/.config/logs/2024.06.14/17.38.58.635304.log
/content/.config/logs/2024.06.14/17.38.14.830445.log
/content/.config/logs/2024.06.14/17.38.38.781607.log
/content/.config/logs/2024.06.14/17.39.09.707862.log
/content/.config/logs/2024.06.14/17.39.10.317752.log
/content/.config/configurations/config_default
/content/sample_data/anscombe.json
/content/sample_data/README.md
/content/sample_data/california_housing_train.csv
/content/sample_data/california_housing_test.csv
/content/sample_data/mnist_train_small.csv
/content/sample_data/mnist_test.csv
In [4]:
import numpy as np # linear algebra
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import plotly.express as px

import warnings
warnings.filterwarnings("ignore")
In [5]:
ass = pd.read_csv('/assessments.csv')
courses = pd.read_csv('/courses.csv')
results = pd.read_csv('/studentAssessment.csv')
info = pd.read_csv('/studentInfo.csv')
reg = pd.read_csv('/studentRegistration.csv')

vle = pd.read_csv('/vle.csv')
In [6]:
svle= pd.read_csv('/studentVle.csv')

a=svle.head(5)
print(a)


b=svle.shape
print(b)
print(svle.dtypes)
svle.describe()

  code_module code_presentation  id_student   id_site  date  sum_click
0         AAA             2013J     28400.0  546652.0 -10.0        4.0
1         AAA             2013J     28400.0  546652.0 -10.0        1.0
2         AAA             2013J     28400.0  546652.0 -10.0        1.0
3         AAA             2013J     28400.0  546614.0 -10.0       11.0
4         AAA             2013J     28400.0  546714.0 -10.0        1.0
(2243524, 6)
code_module           object
code_presentation     object
id_student           float64
id_site              float64
date                 float64
sum_click            float64
dtype: object
Out[6]:
id_student id_site date sum_click
count 2.243523e+06 2.243523e+06 2.243523e+06 2.243523e+06
mean 7.577006e+05 7.291923e+05 8.888535e+01 3.422300e+00
std 6.468020e+05 1.344491e+05 7.374637e+01 6.892611e+00
min 6.516000e+03 5.427920e+05 -2.400000e+01 1.000000e+00
25% 4.785500e+05 5.468790e+05 2.300000e+01 1.000000e+00
50% 5.756810e+05 7.296780e+05 7.300000e+01 2.000000e+00
75% 6.443620e+05 8.770260e+05 1.470000e+02 4.000000e+00
max 2.698588e+06 1.010323e+06 2.690000e+02 2.664000e+03

The table comprises of the following columns:

1.id_site – an identification number of the material. 2.code_module – an identification code for module. 3.code_presentation - the identification code of presentation. 4.activity_type – the role associated with the module material. 5.week_from – the week from which the material is planned to be used. 6.week_to – week until which the material is planned to be used.

In [7]:
null_counts = svle.isnull().sum()
null_counts
Out[7]:
code_module          0
code_presentation    1
id_student           1
id_site              1
date                 1
sum_click            1
dtype: int64
In [8]:
a=info.head(5)
print(a)


b=info.shape
print(b)
print(info.dtypes)
info.describe()

  code_module code_presentation  id_student gender                region  \
0         AAA             2013J       11391      M   East Anglian Region   
1         AAA             2013J       28400      F              Scotland   
2         AAA             2013J       30268      F  North Western Region   
3         AAA             2013J       31604      F     South East Region   
4         AAA             2013J       32885      F  West Midlands Region   

       highest_education imd_band age_band  num_of_prev_attempts  \
0       HE Qualification  90-100%     55<=                     0   
1       HE Qualification   20-30%    35-55                     0   
2  A Level or Equivalent   30-40%    35-55                     0   
3  A Level or Equivalent   50-60%    35-55                     0   
4     Lower Than A Level   50-60%     0-35                     0   

   studied_credits disability final_result  
0              240          N         Pass  
1               60          N         Pass  
2               60          Y    Withdrawn  
3               60          N         Pass  
4               60          N         Pass  
(32593, 12)
code_module             object
code_presentation       object
id_student               int64
gender                  object
region                  object
highest_education       object
imd_band                object
age_band                object
num_of_prev_attempts     int64
studied_credits          int64
disability              object
final_result            object
dtype: object
Out[8]:
id_student num_of_prev_attempts studied_credits
count 3.259300e+04 32593.000000 32593.000000
mean 7.066877e+05 0.163225 79.758691
std 5.491673e+05 0.479758 41.071900
min 3.733000e+03 0.000000 30.000000
25% 5.085730e+05 0.000000 60.000000
50% 5.903100e+05 0.000000 60.000000
75% 6.444530e+05 0.000000 120.000000
max 2.716795e+06 6.000000 655.000000

This file contains demographic information about the students together with their results. File contains the following columns:

1.code_module – an identification code for a module on which the student is registered. 2.code_presentation - the identification code of the presentation during which the student is registered on the module. 3.id_student – a unique identification number for the student. 4.gender – the student’s gender. 5.region – identifies the geographic region, where the student lived while taking the module-presentation. 6.highest_education – highest student education level on entry to the module presentation. 7.imd_band – specifies the Index of Multiple Depravation band of the place where the student lived during the module-presentation. 8.age_band – band of the student’s age. 9.num_of_prev_attempts – the number times the student has attempted this module. 10.studied_credits – the total number of credits for the modules the student is currently studying. 11.disability – indicates whether the student has declared a disability. 12.final_result – student’s final result in the module-presentation.

In [9]:
null_counts = info.isnull().sum()
null_counts
Out[9]:
code_module                0
code_presentation          0
id_student                 0
gender                     0
region                     0
highest_education          0
imd_band                1111
age_band                   0
num_of_prev_attempts       0
studied_credits            0
disability                 0
final_result               0
dtype: int64
In [10]:
stu_info = info['imd_band'].mode()[0]
info['imd_band'].fillna(stu_info, inplace=True)
In [11]:
null_counts = info.isnull().sum()
null_counts
Out[11]:
code_module             0
code_presentation       0
id_student              0
gender                  0
region                  0
highest_education       0
imd_band                0
age_band                0
num_of_prev_attempts    0
studied_credits         0
disability              0
final_result            0
dtype: int64
In [12]:
a=courses.head(5)
print(a)


b=courses.shape
print(b)
print(courses.dtypes)
courses.describe()

  code_module code_presentation  module_presentation_length
0         AAA             2013J                         268
1         AAA             2014J                         269
2         BBB             2013J                         268
3         BBB             2014J                         262
4         BBB             2013B                         240
(22, 3)
code_module                   object
code_presentation             object
module_presentation_length     int64
dtype: object
Out[12]:
module_presentation_length
count 22.000000
mean 255.545455
std 13.654677
min 234.000000
25% 241.000000
50% 261.500000
75% 268.000000
max 269.000000

File contains the list of all available modules and their details. The columns are:

1.code_module – code name of the module, which serves as the identifier.

2.code_presentation – code name of the presentation. It consists of the year and “B” for the presentation starting in February and “J” for the presentation starting in October.

3.length - length of the module-presentation in days.

In [13]:
null_counts = courses.isnull().sum()
null_counts
Out[13]:
code_module                   0
code_presentation             0
module_presentation_length    0
dtype: int64
In [14]:
a=vle.head(5)
print(a)


b=vle.shape
print(b)
print(vle.dtypes)
vle.describe()

   id_site code_module code_presentation activity_type  week_from  week_to
0   546943         AAA             2013J      resource        NaN      NaN
1   546712         AAA             2013J     oucontent        NaN      NaN
2   546998         AAA             2013J      resource        NaN      NaN
3   546888         AAA             2013J           url        NaN      NaN
4   547035         AAA             2013J      resource        NaN      NaN
(6364, 6)
id_site                int64
code_module           object
code_presentation     object
activity_type         object
week_from            float64
week_to              float64
dtype: object
Out[14]:
id_site week_from week_to
count 6.364000e+03 1121.000000 1121.000000
mean 7.260991e+05 15.204282 15.214987
std 1.283151e+05 8.792865 8.779806
min 5.267210e+05 0.000000 0.000000
25% 6.615928e+05 8.000000 8.000000
50% 7.300965e+05 15.000000 15.000000
75% 8.140162e+05 22.000000 22.000000
max 1.077905e+06 29.000000 29.000000

The studentVle.csv file contains information about each student’s interactions with the materials in the VLE. This file contains the following columns:

1.code_module – an identification code for a module. 2.code_presentation - the identification code of the module presentation. 3.id_student – a unique identification number for the student. 4.id_site - an identification number for the VLE material. 5.date – the date of student’s interaction with the material measured as the number of days since the start of the module-presentation. 6.sum_click – the number of times a student interacts with the material in

In [15]:
null_counts = vle.isnull().sum()
null_counts
Out[15]:
id_site                 0
code_module             0
code_presentation       0
activity_type           0
week_from            5243
week_to              5243
dtype: int64

Drop column week_from and week_to

In [16]:
vle=vle.drop(['week_from', 'week_to'], axis=1)
vle.head()
Out[16]:
id_site code_module code_presentation activity_type
0 546943 AAA 2013J resource
1 546712 AAA 2013J oucontent
2 546998 AAA 2013J resource
3 546888 AAA 2013J url
4 547035 AAA 2013J resource
In [17]:
a=ass.head(5)
print(a)


b=ass.shape
print(b)
print(ass.dtypes)
ass.describe()

  code_module code_presentation  id_assessment assessment_type   date  weight
0         AAA             2013J           1752             TMA   19.0    10.0
1         AAA             2013J           1753             TMA   54.0    20.0
2         AAA             2013J           1754             TMA  117.0    20.0
3         AAA             2013J           1755             TMA  166.0    20.0
4         AAA             2013J           1756             TMA  215.0    30.0
(206, 6)
code_module           object
code_presentation     object
id_assessment          int64
assessment_type       object
date                 float64
weight               float64
dtype: object
Out[17]:
id_assessment date weight
count 206.000000 195.000000 206.000000
mean 26473.975728 145.005128 20.873786
std 10098.625521 76.001119 30.384224
min 1752.000000 12.000000 0.000000
25% 15023.250000 71.000000 0.000000
50% 25364.500000 152.000000 12.500000
75% 34891.750000 222.000000 24.250000
max 40088.000000 261.000000 100.000000

This file contains information about assessments in module-presentations. Usually, every presentation has a number of assessments followed by the final exam. CSV contains columns:

1.code_module – identification code of the module, to which the assessment belongs.

2.code_presentation - identification code of the presentation, to which the assessment belongs.

3.id_assessment – identification number of the assessment.

4.assessment_type – type of assessment. Three types of assessments exist: Tutor Marked Assessment (TMA), Computer Marked Assessment (CMA) and Final Exam (Exam).

5.date – information about the final submission date of the assessment calculated as the number of days since the start of the module-presentation. The starting date of the presentation has number 0 (zero).

6.weight - weight of the assessment in %. Typically, Exams are treated separately and have the weight 100%; the sum of all other assessments is 100%. If the information about the final exam date is missing, it is at the end of the last presentation week.

In [18]:
null_counts = ass.isnull().sum()
null_counts
Out[18]:
code_module           0
code_presentation     0
id_assessment         0
assessment_type       0
date                 11
weight                0
dtype: int64
In [19]:
ass = ass.fillna(ass.min())
null_counts = ass.isnull().sum()
null_counts
Out[19]:
code_module          0
code_presentation    0
id_assessment        0
assessment_type      0
date                 0
weight               0
dtype: int64
In [20]:
a=reg.head(5)
print(a)


b=reg.shape
print(b)
print(reg.dtypes)
reg.describe()

  code_module code_presentation  id_student  date_registration  \
0         AAA             2013J       11391             -159.0   
1         AAA             2013J       28400              -53.0   
2         AAA             2013J       30268              -92.0   
3         AAA             2013J       31604              -52.0   
4         AAA             2013J       32885             -176.0   

   date_unregistration  
0                  NaN  
1                  NaN  
2                 12.0  
3                  NaN  
4                  NaN  
(32593, 5)
code_module             object
code_presentation       object
id_student               int64
date_registration      float64
date_unregistration    float64
dtype: object
Out[20]:
id_student date_registration date_unregistration
count 3.259300e+04 32548.000000 10072.000000
mean 7.066877e+05 -69.411300 49.757645
std 5.491673e+05 49.260522 82.460890
min 3.733000e+03 -322.000000 -365.000000
25% 5.085730e+05 -100.000000 -2.000000
50% 5.903100e+05 -57.000000 27.000000
75% 6.444530e+05 -29.000000 109.000000
max 2.716795e+06 167.000000 444.000000

File contains five columns:

1.code_module – an identification code for a module.

2.code_presentation - the identification code of the presentation.

3.id_student – a unique identification number for the student.

4.date_registration – the date of student’s registration on the module presentation, this is the number of days measured relative to the start of the module-presentation (e.g. the negative value -30 means that the student registered to module presentation 30 days before it started).

5.date_unregistration – date of student un-registration from the module presentation, this is the number of days measured relative to the start of the module-presentation.

In [21]:
null_counts = reg.isnull().sum()
null_counts
Out[21]:
code_module                0
code_presentation          0
id_student                 0
date_registration         45
date_unregistration    22521
dtype: int64

almost 70 % data of date_unregistration is missing. it means 70 % students attempts for final exam

In [22]:
# Fill null values in 'date_registration' with the mode of the column
date_registration_mode = reg['date_registration'].mode()[0]
stu_reg=reg['date_registration'].fillna(date_registration_mode, inplace=True)

# Fill null values in 'date_unregistration' with the mode of the column
date_unregistration_mode = reg['date_unregistration'].mode()[0]
reg['date_unregistration'].fillna(date_unregistration_mode, inplace=True)

null_counts = reg.isnull().sum()
null_counts
Out[22]:
code_module            0
code_presentation      0
id_student             0
date_registration      0
date_unregistration    0
dtype: int64
In [23]:
a=results.head(5)
print(a)


b=results.shape
print(b)
print(results.dtypes)
results.describe()

   id_assessment  id_student  date_submitted  is_banked  score
0           1752       11391              18          0   78.0
1           1752       28400              22          0   70.0
2           1752       31604              17          0   72.0
3           1752       32885              26          0   69.0
4           1752       38053              19          0   79.0
(173912, 5)
id_assessment       int64
id_student          int64
date_submitted      int64
is_banked           int64
score             float64
dtype: object
Out[23]:
id_assessment id_student date_submitted is_banked score
count 173912.000000 1.739120e+05 173912.000000 173912.000000 173739.000000
mean 26553.803556 7.051507e+05 116.032942 0.010977 75.799573
std 8829.784254 5.523952e+05 71.484148 0.104194 18.798107
min 1752.000000 6.516000e+03 -11.000000 0.000000 0.000000
25% 15022.000000 5.044290e+05 51.000000 0.000000 65.000000
50% 25359.000000 5.852080e+05 116.000000 0.000000 80.000000
75% 34883.000000 6.344980e+05 173.000000 0.000000 90.000000
max 37443.000000 2.698588e+06 608.000000 1.000000 100.000000

This file contains the following columns:

1.id_assessment – the identification number of the assessment.

2.id_student – a unique identification number for the student.

3.date_submitted – the date of student submission, measured as the number of days since the start of the module presentation.

4.is_banked – a status flag indicating that the assessment result has been transferred from a previous presentation.

5.score – the student’s score in this assessment. The range is from 0 to 100. The score lower than 40 is interpreted as Fail. The marks are in the range from 0 to 100.

In [24]:
null_counts = results.isnull().sum()
null_counts
Out[24]:
id_assessment       0
id_student          0
date_submitted      0
is_banked           0
score             173
dtype: int64
In [25]:
results = results.fillna(results.median())
null_counts = results.isnull().sum()
null_counts
Out[25]:
id_assessment     0
id_student        0
date_submitted    0
is_banked         0
score             0
dtype: int64
In [26]:
# Merge with an inner join
STU_VLE = pd.merge(vle, svle, on=['code_module', 'code_presentation', 'id_site'], how='inner')

Exploratory Data analysis

  1. univariate data analysis
In [27]:
STU_VLE.head()
Out[27]:
id_site code_module code_presentation activity_type id_student date sum_click
0 546943 AAA 2013J resource 75091.0 -10.0 1.0
1 546943 AAA 2013J resource 186149.0 -10.0 1.0
2 546943 AAA 2013J resource 205350.0 -10.0 2.0
3 546943 AAA 2013J resource 1626710.0 -9.0 1.0
4 546943 AAA 2013J resource 2643002.0 -8.0 1.0
In [28]:
STU_VLE\
.groupby(['code_module', 'code_presentation', 'id_student'])\
.agg(total_click = ("sum_click",sum)).reset_index()
Out[28]:
code_module code_presentation id_student total_click
0 AAA 2013J 11391.0 934.0
1 AAA 2013J 28400.0 1435.0
2 AAA 2013J 30268.0 281.0
3 AAA 2013J 31604.0 2158.0
4 AAA 2013J 32885.0 1034.0
... ... ... ... ...
9032 CCC 2014B 2686074.0 34.0
9033 CCC 2014B 2686578.0 779.0
9034 CCC 2014B 2692327.0 1627.0
9035 CCC 2014B 2697181.0 1330.0
9036 CCC 2014B 2698535.0 584.0

9037 rows × 4 columns

In [29]:
total_click_per_student = STU_VLE\
.groupby(['code_module', 'code_presentation', 'id_student'])\
.agg(total_click = ("sum_click",sum))\
.reset_index()

total_click_per_student.head(7)
Out[29]:
code_module code_presentation id_student total_click
0 AAA 2013J 11391.0 934.0
1 AAA 2013J 28400.0 1435.0
2 AAA 2013J 30268.0 281.0
3 AAA 2013J 31604.0 2158.0
4 AAA 2013J 32885.0 1034.0
5 AAA 2013J 38053.0 2445.0
6 AAA 2013J 45462.0 1492.0

student registration info , courses and student info.

In [30]:
regCourses = pd.merge(reg, courses, on=['code_module', 'code_presentation'], how='inner')
regCourses.head()
Out[30]:
code_module code_presentation id_student date_registration date_unregistration module_presentation_length
0 AAA 2013J 11391 -159.0 12.0 268
1 AAA 2013J 28400 -53.0 12.0 268
2 AAA 2013J 30268 -92.0 12.0 268
3 AAA 2013J 31604 -52.0 12.0 268
4 AAA 2013J 32885 -176.0 12.0 268
In [31]:
# @title code_module vs code_presentation

from matplotlib import pyplot as plt
import seaborn as sns
import pandas as pd
plt.subplots(figsize=(8, 8))
df_2dhist = pd.DataFrame({
    x_label: grp['code_presentation'].value_counts()
    for x_label, grp in regCourses.groupby('code_module')
})
sns.heatmap(df_2dhist, cmap='viridis')
plt.xlabel('code_module')
_ = plt.ylabel('code_presentation')
In [32]:
# @title code_presentation

from matplotlib import pyplot as plt
import seaborn as sns
regCourses.groupby('code_presentation').size().plot(kind='barh', color=sns.palettes.mpl_palette('Dark2'))
plt.gca().spines[['top', 'right',]].set_visible(False)
In [33]:
# @title code_module

from matplotlib import pyplot as plt
import seaborn as sns
regCourses.groupby('code_module').size().plot(kind='barh', color=sns.palettes.mpl_palette('Dark2'))
plt.gca().spines[['top', 'right',]].set_visible(False)
In [34]:
regCoursesInfo = pd.merge(regCourses, info, on=['code_module', 'code_presentation', 'id_student'], how='inner')
regCoursesInfo.head()
Out[34]:
code_module code_presentation id_student date_registration date_unregistration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability final_result
0 AAA 2013J 11391 -159.0 12.0 268 M East Anglian Region HE Qualification 90-100% 55<= 0 240 N Pass
1 AAA 2013J 28400 -53.0 12.0 268 F Scotland HE Qualification 20-30% 35-55 0 60 N Pass
2 AAA 2013J 30268 -92.0 12.0 268 F North Western Region A Level or Equivalent 30-40% 35-55 0 60 Y Withdrawn
3 AAA 2013J 31604 -52.0 12.0 268 F South East Region A Level or Equivalent 50-60% 35-55 0 60 N Pass
4 AAA 2013J 32885 -176.0 12.0 268 F West Midlands Region Lower Than A Level 50-60% 0-35 0 60 N Pass
In [35]:
# @title gender

from matplotlib import pyplot as plt
import seaborn as sns
regCoursesInfo.groupby('gender').size().plot(kind='barh', color=sns.palettes.mpl_palette('Dark2'))
plt.gca().spines[['top', 'right',]].set_visible(False)
In [36]:
# @title highest_education

from matplotlib import pyplot as plt
import seaborn as sns
regCoursesInfo.groupby('highest_education').size().plot(kind='barh', color=sns.palettes.mpl_palette('Dark2'))
plt.gca().spines[['top', 'right',]].set_visible(False)
In [37]:
# @title date_registration vs date_unregistration

from matplotlib import pyplot as plt
regCoursesInfo.plot(kind='scatter', x='date_registration', y='date_unregistration', s=32, alpha=.8)
plt.gca().spines[['top', 'right',]].set_visible(False)

Assessment info and assessment results

In [38]:
assResults = pd.merge(ass, results, on=['id_assessment'], how='inner')

# Rearrange column names

assResults = assResults[['id_student', 'code_module', 'code_presentation', 'id_assessment', 'assessment_type', 'date', 'date_submitted', 'weight', 'is_banked', 'score']]

assResults.head()
Out[38]:
id_student code_module code_presentation id_assessment assessment_type date date_submitted weight is_banked score
0 11391 AAA 2013J 1752 TMA 19.0 18 10.0 0 78.0
1 28400 AAA 2013J 1752 TMA 19.0 22 10.0 0 70.0
2 31604 AAA 2013J 1752 TMA 19.0 17 10.0 0 72.0
3 32885 AAA 2013J 1752 TMA 19.0 26 10.0 0 69.0
4 38053 AAA 2013J 1752 TMA 19.0 19 10.0 0 79.0

Feature Engineering

How it will be calculated: To calculate the total weight of all modules, we need to remember that most final exams are missing from the Results table. Multiply the weight of the assignment with its score. Aggregate the dataframe per weightscore per module per module presentation with the sum function. Calculate total recorded weight of module (recorded total is key here as most modules are missing their final exam). Now calculate weighted scores - divide summed weightscore by total recorded weight of module

In [39]:
scores=assResults
scores[scores['assessment_type'] == 'Exam'][['code_module', 'code_presentation', 'id_assessment']]\
.groupby(['code_module', 'code_presentation'])\
.nunique()
Out[39]:
id_assessment
code_module code_presentation
CCC 2014B 1
2014J 1
DDD 2013B 1
2013J 1
2014B 1
2014J 1

CCC module only has results for 1 exam when the module should have 2 exams in total.

DDD module has results for the final exam (DDD module should have one exam in total).

In [40]:
### Make helper columns ###
# (a) Add column multiplying weight and score
scores['weight*score'] = scores['weight']*scores['score']
# (b) Aggregate recorded weight*score per student
    # per module presentation
sum_scores = scores\
.groupby(['id_student', 'code_module', 'code_presentation'])\
.agg(weightByScore = ('weight*score', sum))\
.reset_index()
# (c) Calculate total recorded weight of module
# (c.i) Get total weight of modules
total_weight = ass\
.groupby(['code_module', 'code_presentation'])\
.agg(total_weight = ('weight', sum))\
.reset_index()
# (c.ii) Subtract 100 to account for missing exams
total_weight['total_weight'] = total_weight['total_weight']-100
# (c.iii) Mark module DDD as having 200 credits
total_weight.loc[(total_weight.code_module == 'DDD'), 'total_weight'] = 200

### Calculate weighted score ###
# (a) Merge sum_scores and total_weight tables
score_weights = pd.merge(sum_scores, total_weight, on=['code_module', 'code_presentation'], how='inner')
# (b) Calculate weighted score
score_weights['weighted_score'] = score_weights['weightByScore'] / score_weights['total_weight']
# (c) Drop helper columns
score_weights.drop(columns=['weightByScore', 'total_weight'], inplace=True)
In [41]:
score_weights.head()
Out[41]:
id_student code_module code_presentation weighted_score
0 6516 AAA 2014J 63.5
1 24734 AAA 2014J 47.5
2 26192 AAA 2014J 86.3
3 28061 AAA 2014J 77.4
4 31600 AAA 2014J 55.5

Late submission

How will be calculated Calculate the difference between the deadline and the actual submission date.

Make a new column - if the difference between dates is more that ), the submission was late.

Aggregate by student ID, module, and module presenation

In [42]:
# Calculate the difference between the submission dates
lateSubmission = assResults.assign(submission_days=assResults['date_submitted']-assResults['date'])
# Make a column indicating if the submission was late or not
lateSubmission = lateSubmission.assign(late_submission=lateSubmission['submission_days'] > 0)

lateSubmission.head()
Out[42]:
id_student code_module code_presentation id_assessment assessment_type date date_submitted weight is_banked score weight*score submission_days late_submission
0 11391 AAA 2013J 1752 TMA 19.0 18 10.0 0 78.0 780.0 -1.0 False
1 28400 AAA 2013J 1752 TMA 19.0 22 10.0 0 70.0 700.0 3.0 True
2 31604 AAA 2013J 1752 TMA 19.0 17 10.0 0 72.0 720.0 -2.0 False
3 32885 AAA 2013J 1752 TMA 19.0 26 10.0 0 69.0 690.0 7.0 True
4 38053 AAA 2013J 1752 TMA 19.0 19 10.0 0 79.0 790.0 0.0 False
In [43]:
lateSubmission[(lateSubmission['assessment_type'] == 'Exam') & (lateSubmission['late_submission'] == True)]
Out[43]:
id_student code_module code_presentation id_assessment assessment_type date date_submitted weight is_banked score weight*score submission_days late_submission
52923 558914 CCC 2014B 24290 Exam 12.0 230 100.0 0 32.0 3200.0 218.0 True
52924 559706 CCC 2014B 24290 Exam 12.0 234 100.0 0 78.0 7800.0 222.0 True
52925 559770 CCC 2014B 24290 Exam 12.0 230 100.0 0 54.0 5400.0 218.0 True
52926 560114 CCC 2014B 24290 Exam 12.0 230 100.0 0 64.0 6400.0 218.0 True
52927 560311 CCC 2014B 24290 Exam 12.0 234 100.0 0 100.0 10000.0 222.0 True
... ... ... ... ... ... ... ... ... ... ... ... ... ...
95980 652680 DDD 2014J 25368 Exam 12.0 243 100.0 0 91.0 9100.0 231.0 True
95981 652732 DDD 2014J 25368 Exam 12.0 249 100.0 0 47.0 4700.0 237.0 True
95982 652965 DDD 2014J 25368 Exam 12.0 242 100.0 0 87.0 8700.0 230.0 True
95983 653051 DDD 2014J 25368 Exam 12.0 242 100.0 0 27.0 2700.0 230.0 True
95984 653319 DDD 2014J 25368 Exam 12.0 250 100.0 0 64.0 6400.0 238.0 True

2872 rows × 13 columns

In [44]:
# Aggregate per student per module presentation
total_late_per_student = lateSubmission\
.groupby(['id_student', 'code_module', 'code_presentation'])\
.agg(total_late_submission = ('late_submission', sum))\
.reset_index()

total_late_per_student.head()
Out[44]:
id_student code_module code_presentation total_late_submission
0 6516 AAA 2014J 0
1 8462 DDD 2013J 1
2 8462 DDD 2014J 0
3 11391 AAA 2013J 0
4 23629 BBB 2013B 3
In [45]:
# Make a df with total number of all assessments per student per module presentation
total_count_assessments = lateSubmission[['id_student', 'code_module', 'code_presentation', 'id_assessment']]\
.groupby(['id_student', 'code_module', 'code_presentation'])\
.size()\
.reset_index(name='total_assessments')

total_count_assessments.head()
Out[45]:
id_student code_module code_presentation total_assessments
0 6516 AAA 2014J 5
1 8462 DDD 2013J 3
2 8462 DDD 2014J 4
3 11391 AAA 2013J 5
4 23629 BBB 2013B 4
In [46]:
# Merge df with total late assessements and total count assessments
late_rate_per_student = pd.merge(total_late_per_student, total_count_assessments, on=['id_student', 'code_module', 'code_presentation'], how='inner')
# Make a new column with late submission rate
late_rate_per_student['late_rate'] = late_rate_per_student['total_late_submission'] / late_rate_per_student['total_assessments']
# Drop helper columns
late_rate_per_student.drop(columns=['total_late_submission', 'total_assessments'], inplace=True)

late_rate_per_student
Out[46]:
id_student code_module code_presentation late_rate
0 6516 AAA 2014J 0.000000
1 8462 DDD 2013J 0.333333
2 8462 DDD 2014J 0.000000
3 11391 AAA 2013J 0.000000
4 23629 BBB 2013B 0.750000
... ... ... ... ...
25838 2698257 AAA 2013J 0.400000
25839 2698535 CCC 2014B 0.750000
25840 2698535 EEE 2013J 0.250000
25841 2698577 BBB 2014J 0.000000
25842 2698588 BBB 2014J 0.200000

25843 rows × 4 columns

Fail rate

In [47]:
# Define function for marking failed assignments
passRate = assResults
passRate = passRate.assign(fail=passRate['score'] < 40)

passRate.head()
Out[47]:
id_student code_module code_presentation id_assessment assessment_type date date_submitted weight is_banked score weight*score fail
0 11391 AAA 2013J 1752 TMA 19.0 18 10.0 0 78.0 780.0 False
1 28400 AAA 2013J 1752 TMA 19.0 22 10.0 0 70.0 700.0 False
2 31604 AAA 2013J 1752 TMA 19.0 17 10.0 0 72.0 720.0 False
3 32885 AAA 2013J 1752 TMA 19.0 26 10.0 0 69.0 690.0 False
4 38053 AAA 2013J 1752 TMA 19.0 19 10.0 0 79.0 790.0 False
In [48]:
# Aggregate per student per module presentation
total_fails_per_student = passRate\
.groupby(['id_student', 'code_module', 'code_presentation'])\
.agg(total_fails = ("fail",sum))\
.reset_index()

total_fails_per_student.head()
Out[48]:
id_student code_module code_presentation total_fails
0 6516 AAA 2014J 0
1 8462 DDD 2013J 0
2 8462 DDD 2014J 0
3 11391 AAA 2013J 0
4 23629 BBB 2013B 0
In [49]:
# Merge df with total fails and total count assessments
fail_rate_per_student = pd.merge(total_fails_per_student, total_count_assessments, on=['id_student', 'code_module', 'code_presentation'], how='inner')
# Make a new column with late submission rate
fail_rate_per_student['fail_rate'] = fail_rate_per_student['total_fails'] / fail_rate_per_student['total_assessments']
# Drop helper columns
fail_rate_per_student.drop(columns=['total_fails', 'total_assessments'], inplace=True)

fail_rate_per_student
Out[49]:
id_student code_module code_presentation fail_rate
0 6516 AAA 2014J 0.00
1 8462 DDD 2013J 0.00
2 8462 DDD 2014J 0.00
3 11391 AAA 2013J 0.00
4 23629 BBB 2013B 0.00
... ... ... ... ...
25838 2698257 AAA 2013J 0.00
25839 2698535 CCC 2014B 0.75
25840 2698535 EEE 2013J 0.25
25841 2698577 BBB 2014J 0.00
25842 2698588 BBB 2014J 0.00

25843 rows × 4 columns

Merged all result tables

In [50]:
assessments = pd.merge(score_weights, late_rate_per_student, on=['id_student', 'code_module', 'code_presentation'], how='inner')
assessments = pd.merge(assessments, fail_rate_per_student, on=['id_student', 'code_module', 'code_presentation'], how='inner')

assessments.head()
Out[50]:
id_student code_module code_presentation weighted_score late_rate fail_rate
0 6516 AAA 2014J 63.5 0.0 0.0
1 24734 AAA 2014J 47.5 0.8 0.2
2 26192 AAA 2014J 86.3 0.0 0.0
3 28061 AAA 2014J 77.4 0.2 0.0
4 31600 AAA 2014J 55.5 0.6 0.0

Merge all tables

In [51]:
merged = pd.merge(regCoursesInfo, total_click_per_student, on=['id_student', 'code_module', 'code_presentation'], how='left')

merged.head()
Out[51]:
code_module code_presentation id_student date_registration date_unregistration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability final_result total_click
0 AAA 2013J 11391 -159.0 12.0 268 M East Anglian Region HE Qualification 90-100% 55<= 0 240 N Pass 934.0
1 AAA 2013J 28400 -53.0 12.0 268 F Scotland HE Qualification 20-30% 35-55 0 60 N Pass 1435.0
2 AAA 2013J 30268 -92.0 12.0 268 F North Western Region A Level or Equivalent 30-40% 35-55 0 60 Y Withdrawn 281.0
3 AAA 2013J 31604 -52.0 12.0 268 F South East Region A Level or Equivalent 50-60% 35-55 0 60 N Pass 2158.0
4 AAA 2013J 32885 -176.0 12.0 268 F West Midlands Region Lower Than A Level 50-60% 0-35 0 60 N Pass 1034.0
In [52]:
merged = pd.merge(merged, assessments, on=['id_student', 'code_module', 'code_presentation'], how='left')

merged.head()
Out[52]:
code_module code_presentation id_student date_registration date_unregistration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability final_result total_click weighted_score late_rate fail_rate
0 AAA 2013J 11391 -159.0 12.0 268 M East Anglian Region HE Qualification 90-100% 55<= 0 240 N Pass 934.0 82.4 0.0 0.0
1 AAA 2013J 28400 -53.0 12.0 268 F Scotland HE Qualification 20-30% 35-55 0 60 N Pass 1435.0 65.4 0.4 0.0
2 AAA 2013J 30268 -92.0 12.0 268 F North Western Region A Level or Equivalent 30-40% 35-55 0 60 Y Withdrawn 281.0 NaN NaN NaN
3 AAA 2013J 31604 -52.0 12.0 268 F South East Region A Level or Equivalent 50-60% 35-55 0 60 N Pass 2158.0 76.3 0.0 0.0
4 AAA 2013J 32885 -176.0 12.0 268 F West Midlands Region Lower Than A Level 50-60% 0-35 0 60 N Pass 1034.0 55.0 1.0 0.4

SPLIT THE DATASET

In [53]:
from sklearn.model_selection import train_test_split

train, test = train_test_split(merged, test_size=0.2, random_state=42, stratify=merged['code_module'])

Final Cleaning

. Missing values

In [54]:
train.isnull().sum()
Out[54]:
code_module                       0
code_presentation                 0
id_student                        0
date_registration                 0
date_unregistration               0
module_presentation_length        0
gender                            0
region                            0
highest_education                 0
imd_band                          0
age_band                          0
num_of_prev_attempts              0
studied_credits                   0
disability                        0
final_result                      0
total_click                   18866
weighted_score                 7117
late_rate                      5434
fail_rate                      5434
dtype: int64

IMD band

In [55]:
train\
[train['imd_band'].isnull()]\
.head()
Out[55]:
code_module code_presentation id_student date_registration date_unregistration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability final_result total_click weighted_score late_rate fail_rate
In [56]:
train.head()
Out[56]:
code_module code_presentation id_student date_registration date_unregistration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability final_result total_click weighted_score late_rate fail_rate
16563 DDD 2014B 464685 -36.0 12.0 241 F East Anglian Region A Level or Equivalent 50-60% 0-35 1 70 N Pass NaN 64.9125 0.142857 0.0
16006 DDD 2013J 606593 -32.0 122.0 261 M Scotland A Level or Equivalent 20-30% 0-35 0 120 N Withdrawn NaN 14.8625 0.000000 0.0
12884 CCC 2014J 1674799 -21.0 12.0 269 M South Region HE Qualification 70-80% 55<= 0 60 N Pass NaN 61.4300 0.875000 0.0
27781 FFF 2014J 195246 -124.0 12.0 269 M West Midlands Region Lower Than A Level 10-20 0-35 0 60 N Fail NaN 39.5000 0.181818 0.0
25125 FFF 2013J 582080 -114.0 12.0 268 M South East Region Lower Than A Level 10-20 35-55 0 60 N Fail NaN 9.2500 0.000000 0.0

How to fill out missing values here - fill out according to the most frequent band for that region:

1.Find which regions have null imd_band values

2.Find what band is the most frequent one for that region

3.Replace null values with most frequent values for that region.

In [57]:
regions_list = list(train[train['imd_band'].isnull()]['region'].unique())

for i in regions_list:

    result = train[train['region'] == i].imd_band.mode()

    print(f'{i} IMD band : \n', result)
In [58]:
# Replace all null values with respective most frequent imd_bands
regions_list = list(train\
                    [train['imd_band'].isnull()]['region']\
                    .unique())

for i in regions_list:
    train['imd_band'] = np.where( ( (train['imd_band'].isnull()) & (train['region'] == i) ),
                                           train[train['region'] == i].imd_band.mode(),
                                           train['imd_band']
                                    )

Date registration

In [59]:
# Make a new dataframe just with rows that have null values for the registration date
reg_date_nulls_in_reg = train\
[train['date_registration'].isnull()]
In [60]:
# What are their final results?
column = reg_date_nulls_in_reg['final_result']

unique, counts = np.unique(column, return_counts = True)

dict(zip(unique, counts))
Out[60]:
{}
In [61]:
# Get median registration date
train.date_registration.median()
Out[61]:
-57.0
In [62]:
# Replace NaN values with date_unreg minus the median (note, the median is negative)
train['date_registration'] = np.where( (train['date_registration'].isnull()),
                                           train['date_unregistration'] + train.date_registration.median(),
                                           train['date_registration']
                                    )
# Replace remaining NaNs with -57
train['date_registration'] = np.where( (train['date_registration'].isnull()),
                                           train.date_registration.median(),
                                           train['date_registration']
                                    )

Total_click

In [63]:
train['total_click'] = train['total_click'].replace(np.nan).fillna(0)

weighted_score

In [64]:
train['weighted_score'] = train['weighted_score'].replace(np.nan).fillna(0)

late_rate

In [65]:
train['late_rate'] = train['late_rate'].replace(np.nan).fillna(1.0)

fail_rate

In [66]:
train['fail_rate'] = train['fail_rate'].replace(np.nan).fillna(1.0)

drop columns

In [67]:
# Make a copy of training and test datasets for classification
train_class = train.copy()
test_class = test.copy()

For Regresion

Date unregistration has been dropped as it should be the same metric as Withdrawal in the final results column of Student Information table. I'm assuming the prediction for who is likely to withdraw would only be useful if trying to predict future withdrawals, trying to predict if a student has withdrawn when we know they have withdrawn is useless (date_unregistration is essentially another target). For the regression problem due to the contuinous nature of the target variable, we can't distinguish between fails and witdrawals, so all withdrawals will be treated as fails (score < 40%).

In [68]:
# Drop unneeded columns
train.drop(columns=['id_student'], inplace=True)
train.drop(columns=['final_result'], inplace=True)
train.drop(columns=['date_unregistration'], inplace=True)

train.head()
Out[68]:
code_module code_presentation date_registration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability total_click weighted_score late_rate fail_rate
16563 DDD 2014B -36.0 241 F East Anglian Region A Level or Equivalent 50-60% 0-35 1 70 N 0.0 64.9125 0.142857 0.0
16006 DDD 2013J -32.0 261 M Scotland A Level or Equivalent 20-30% 0-35 0 120 N 0.0 14.8625 0.000000 0.0
12884 CCC 2014J -21.0 269 M South Region HE Qualification 70-80% 55<= 0 60 N 0.0 61.4300 0.875000 0.0
27781 FFF 2014J -124.0 269 M West Midlands Region Lower Than A Level 10-20 0-35 0 60 N 0.0 39.5000 0.181818 0.0
25125 FFF 2013J -114.0 268 M South East Region Lower Than A Level 10-20 35-55 0 60 N 0.0 9.2500 0.000000 0.0

For Classificatio

In [69]:
# Drop unneeded columns
train_class.drop(columns=['id_student'], inplace=True)
train_class.drop(columns=['date_unregistration'], inplace=True)
# Drop columns on assessments
train_class.drop(columns=['weighted_score'], inplace=True)
train_class.drop(columns=['late_rate'], inplace=True)
train_class.drop(columns=['fail_rate'], inplace=True)


train_class.head()
Out[69]:
code_module code_presentation date_registration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability final_result total_click
16563 DDD 2014B -36.0 241 F East Anglian Region A Level or Equivalent 50-60% 0-35 1 70 N Pass 0.0
16006 DDD 2013J -32.0 261 M Scotland A Level or Equivalent 20-30% 0-35 0 120 N Withdrawn 0.0
12884 CCC 2014J -21.0 269 M South Region HE Qualification 70-80% 55<= 0 60 N Pass 0.0
27781 FFF 2014J -124.0 269 M West Midlands Region Lower Than A Level 10-20 0-35 0 60 N Fail 0.0
25125 FFF 2013J -114.0 268 M South East Region Lower Than A Level 10-20 35-55 0 60 N Fail 0.0

Cleaning test set for regression

In [70]:
'''IMD BAND'''
# Replace all null values with respective most frequent imd_bands
regions_list = list(test\
                    [test['imd_band'].isnull()]['region']\
                    .unique())

for i in regions_list:
    test['imd_band'] = np.where( ( (test['imd_band'].isnull()) & (test['region'] == i) ),
                                           test[test['region'] == i].imd_band.mode(),
                                           test['imd_band']
                                    )

'''DATE REGISTRATION'''
# Get registration date median
reg_date_median = test.date_registration.median()


# Replace NaN values with date_unreg minus 57 days
test['date_registration'] = np.where( (test['date_registration'].isnull()),
                                           test['date_unregistration'] + reg_date_median,
                                           test['date_registration']
                                    )
# Replace remaining NaNs with -57
test['date_registration'] = np.where( (test['date_registration'].isnull()),
                                           reg_date_median,
                                           test['date_registration']
                                    )

'''Rest of null values'''
test['total_click'] = test['total_click'].replace(np.nan).fillna(0)
test['weighted_score'] = test['weighted_score'].replace(np.nan).fillna(0)
test['late_rate'] = test['late_rate'].replace(np.nan).fillna(1.0)
test['fail_rate'] = test['fail_rate'].replace(np.nan).fillna(1.0)

'''Drop unneeded columns'''
# Drop ID, final result, and date unregistration columns
test.drop(columns=['id_student'], inplace=True)
test.drop(columns=['final_result'], inplace=True)
test.drop(columns=['date_unregistration'], inplace=True)

for classification

In [71]:
'''IMD BAND'''
# Replace all null values with respective most frequent imd_bands
regions_list = list(test_class\
                    [test_class['imd_band'].isnull()]['region']\
                    .unique())

for i in regions_list:
    test_class['imd_band'] = np.where( ( (test_class['imd_band'].isnull()) & (test_class['region'] == i) ),
                                           test_class[test_class['region'] == i].imd_band.mode(),
                                           test_class['imd_band']
                                    )

'''DATE REGISTRATION'''
# Get registration date median
reg_date_median = test_class.date_registration.median()


# Replace NaN values with date_unreg minus 57 days
test_class['date_registration'] = np.where( (test_class['date_registration'].isnull()),
                                           test_class['date_unregistration'] + reg_date_median,
                                           test_class['date_registration']
                                    )
# Replace remaining NaNs with -57
test_class['date_registration'] = np.where( (test_class['date_registration'].isnull()),
                                           reg_date_median,
                                           test_class['date_registration']
                                    )

'''Rest of null values'''
test_class['total_click'] = test_class['total_click'].replace(np.nan).fillna(0)

'''Drop unneeded columns'''
# Drop ID, final result, and date unregistration columns
test_class.drop(columns=['id_student'], inplace=True)
test_class.drop(columns=['date_unregistration'], inplace=True)
# Drop columns on assessments
test_class.drop(columns=['weighted_score'], inplace=True)
test_class.drop(columns=['late_rate'], inplace=True)
test_class.drop(columns=['fail_rate'], inplace=True)

Univariate analysis

numerical data

In [72]:
train.describe().transpose()
Out[72]:
count mean std min 25% 50% 75% max
date_registration 26074.0 -69.428703 49.313015 -322.0 -101.00 -57.00 -29.000000 167.0
module_presentation_length 26074.0 256.033673 13.168106 234.0 241.00 262.00 268.000000 269.0
num_of_prev_attempts 26074.0 0.162307 0.480874 0.0 0.00 0.00 0.000000 6.0
studied_credits 26074.0 79.730958 40.841449 30.0 60.00 60.00 120.000000 540.0
total_click 26074.0 841.368375 1184.959853 0.0 187.00 470.00 1053.000000 16440.0
weighted_score 26074.0 51.694506 30.270418 0.0 19.75 59.45 77.667500 100.0
late_rate 26074.0 0.313873 0.328966 0.0 0.00 0.20 0.555556 1.0
fail_rate 26074.0 0.068107 0.185976 0.0 0.00 0.00 0.000000 1.0

Distribution Plots

In [73]:
# Create statistics summaries with skew, mean, and median
# Produce a dataframe with just numerical columns
df_num = train.select_dtypes(include=np.number)

for col in df_num.columns:

    skew = df_num[col].skew()
    mean = df_num[col].mean()
    median = df_num[col].median()
    print(f'\tSummary for {col.upper()}')
    print(f'Skewness of {col}\t: {skew}')
    print(f'Mean {col} :\t {mean}')
    print(f'Median {col} :\t {median} \n')

	Summary for DATE_REGISTRATION
Skewness of date_registration	: -1.0048926915675045
Mean date_registration :	 -69.42870292245148
Median date_registration :	 -57.0 

	Summary for MODULE_PRESENTATION_LENGTH
Skewness of module_presentation_length	: -0.40871442091912236
Mean module_presentation_length :	 256.0336733911176
Median module_presentation_length :	 262.0 

	Summary for NUM_OF_PREV_ATTEMPTS
Skewness of num_of_prev_attempts	: 3.867098396716384
Mean num_of_prev_attempts :	 0.16230727928204342
Median num_of_prev_attempts :	 0.0 

	Summary for STUDIED_CREDITS
Skewness of studied_credits	: 1.7319757093939532
Mean studied_credits :	 79.73095804249444
Median studied_credits :	 60.0 

	Summary for TOTAL_CLICK
Skewness of total_click	: 4.23905567116546
Mean total_click :	 841.3683746260642
Median total_click :	 470.0 

	Summary for WEIGHTED_SCORE
Skewness of weighted_score	: -0.37197197116891306
Mean weighted_score :	 51.69450554191915
Median weighted_score :	 59.45 

	Summary for LATE_RATE
Skewness of late_rate	: 0.6792572546548434
Mean late_rate :	 0.31387261807508027
Median late_rate :	 0.2 

	Summary for FAIL_RATE
Skewness of fail_rate	: 3.5859735248273523
Mean fail_rate :	 0.06810720036442978
Median fail_rate :	 0.0
In [74]:
train.hist(bins=50, figsize=(20,15))
plt.show()

Target Variable - weighted_score

In [75]:
from plotnine import ggplot, aes, geom_boxplot, ggtitle, coord_flip

(
    ggplot(train)
    + aes(x=0, y='weighted_score')
    + geom_boxplot(outlier_color='crimson')
    + ggtitle("Distribution of scores per module")
    + coord_flip()
)
Out[75]:
<Figure Size: (640 x 480)>

We can see the target variable has two peaks and is not normally distributed, but it doesn't have any outliers. We may wish to transform the target at some point to improve our models. This notebook shows very basic analysis though, so we will not be doing this, but it's something to keep in mind when using certain models (like the linear regression that assumes the distributions are normal).

Correlation Matrix

In [76]:
# Let's make a correlation heatmap
plt.figure(figsize=(10,10))
sns.heatmap(df_num.corr(), annot=True, cmap="coolwarm", );

Univariate analysis : Categorical data

In [77]:
train.head()
Out[77]:
code_module code_presentation date_registration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability total_click weighted_score late_rate fail_rate
16563 DDD 2014B -36.0 241 F East Anglian Region A Level or Equivalent 50-60% 0-35 1 70 N 0.0 64.9125 0.142857 0.0
16006 DDD 2013J -32.0 261 M Scotland A Level or Equivalent 20-30% 0-35 0 120 N 0.0 14.8625 0.000000 0.0
12884 CCC 2014J -21.0 269 M South Region HE Qualification 70-80% 55<= 0 60 N 0.0 61.4300 0.875000 0.0
27781 FFF 2014J -124.0 269 M West Midlands Region Lower Than A Level 10-20 0-35 0 60 N 0.0 39.5000 0.181818 0.0
25125 FFF 2013J -114.0 268 M South East Region Lower Than A Level 10-20 35-55 0 60 N 0.0 9.2500 0.000000 0.0
In [78]:
numerical_columns = ['weighted_score',  'total_click'  , 'module_presentation_length', 'date_registration',
                       'studied_credits', 'num_of_prev_attempts', 'late_rate', 'fail_rate'
]


df_num = train[numerical_columns]

# Print the DataFrame with categorical columns
print(df_num.head())

# Select specific categorical columns
categorical_columns = [
    'code_module', 'code_presentation', 'gender', 'region', 'imd_band',
    'highest_education', 'age_band', 'disability'
]

df_cat = train[categorical_columns]

# Print the DataFrame with categorical columns
df_cat.head()

       weighted_score  total_click  module_presentation_length  \
16563         64.9125          0.0                         241   
16006         14.8625          0.0                         261   
12884         61.4300          0.0                         269   
27781         39.5000          0.0                         269   
25125          9.2500          0.0                         268   

       date_registration  studied_credits  num_of_prev_attempts  late_rate  \
16563              -36.0               70                     1   0.142857   
16006              -32.0              120                     0   0.000000   
12884              -21.0               60                     0   0.875000   
27781             -124.0               60                     0   0.181818   
25125             -114.0               60                     0   0.000000   

       fail_rate  
16563        0.0  
16006        0.0  
12884        0.0  
27781        0.0  
25125        0.0
Out[78]:
code_module code_presentation gender region imd_band highest_education age_band disability
16563 DDD 2014B F East Anglian Region 50-60% A Level or Equivalent 0-35 N
16006 DDD 2013J M Scotland 20-30% A Level or Equivalent 0-35 N
12884 CCC 2014J M South Region 70-80% HE Qualification 55<= N
27781 FFF 2014J M West Midlands Region 10-20 Lower Than A Level 0-35 N
25125 FFF 2013J M South East Region 10-20 Lower Than A Level 35-55 N
In [79]:
df_num\
.drop(columns=['weighted_score'])\
.corrwith(train['weighted_score']).plot.bar(
        figsize = (6, 4), title = "Correlation with Target", fontsize = 12,
        rot = 90, grid = True);
In [80]:
df_num.corrwith(train['weighted_score']).sort_values(ascending=False)
Out[80]:
weighted_score                1.000000
total_click                   0.114282
module_presentation_length    0.048314
date_registration            -0.003783
studied_credits              -0.061500
num_of_prev_attempts         -0.088768
late_rate                    -0.241835
fail_rate                    -0.358006
dtype: float64

categorical data

In [81]:
df_cat.head()
Out[81]:
code_module code_presentation gender region imd_band highest_education age_band disability
16563 DDD 2014B F East Anglian Region 50-60% A Level or Equivalent 0-35 N
16006 DDD 2013J M Scotland 20-30% A Level or Equivalent 0-35 N
12884 CCC 2014J M South Region 70-80% HE Qualification 55<= N
27781 FFF 2014J M West Midlands Region 10-20 Lower Than A Level 0-35 N
25125 FFF 2013J M South East Region 10-20 Lower Than A Level 35-55 N
In [82]:
# Set the plot number for the first subplot function
plot_number = 1

# Set sizes for all plots
plt.figure(figsize=(15, 15)) # create a figure object
plt.subplots_adjust(hspace = 0.5) # set the size of subplots

for col in df_cat[['code_module', 'code_presentation', 'gender', 'region']]:
    # Call countplot on each column
    plt.subplot(4, 2, plot_number)
    sns.countplot(
        y=col,
        data=df_cat,
        order=df_cat[col].value_counts().index
    )
    plt.title(f'{col.capitalize()} Countplot')
    plt.xlabel('')
    plt.ylabel('')

    plot_number = plot_number + 1 # set a new plot number for the next subplot function
    # Add relative frequency labels:
    n_points = df_cat.shape[0]
    col_counts = df_cat[col].value_counts()
    locs, labels = plt.yticks()   # get the current tick locations and labels

    # loop through each pair of locations and labels
    for loc, label in zip(locs, labels):

        # get the text property for the label to get the correct count
        count = col_counts[label.get_text()]
        pct_string = '{:0.1f}%'.format(100*count/n_points)

        # print the annotation at the top of the bar
        plt.text(x=count, y=loc, s=pct_string, ha='left', va='center', color='k')
plt.tight_layout()
In [83]:
# Set the plot number for the first subplot function
plot_number = 1

# Set sizes for all plots
plt.figure(figsize=(15, 15)) # create a figure object
plt.subplots_adjust(hspace = 0.5) # set the size of subplots

for col in df_cat[['highest_education','imd_band', 'age_band', 'disability']]:
    # Call countplot on each column
    plt.subplot(4, 2, plot_number)
    sns.countplot(
        y=col,
        data=df_cat,
        order=df_cat[col].value_counts().index
    )
    plt.title(f'{col.capitalize()} Countplot')
    plt.xlabel('')
    plt.ylabel('')

    plot_number = plot_number + 1 # set a new plot number for the next subplot function
    # Add relative frequency labels:
    n_points = df_cat.shape[0]
    col_counts = df_cat[col].value_counts()
    locs, labels = plt.yticks()   # get the current tick locations and labels

    # loop through each pair of locations and labels
    for loc, label in zip(locs, labels):

        # get the text property for the label to get the correct count
        count = col_counts[label.get_text()]
        pct_string = '{:0.1f}%'.format(100*count/n_points)

        # print the annotation at the top of the bar
        plt.text(x=count, y=loc, s=pct_string, ha='left', va='center', color='k')
plt.tight_layout()

Very few students with no formal education (1%) Very few students with post-grad qualifications. These two categories should be merged with 'Lower Than A Level' and 'HE Qualification', respectivelly, as with so little data these two categories are not likely to bring much insight.

Change education categories

In [84]:
# Rename 'no formal quals' into 'lower than a level'
train['highest_education'] = np.where( (train['highest_education'] == 'No Formal quals'),
                                           'Lower Than A Level',
                                           train['highest_education']
                                    )

# Rename post-grads
train['highest_education'] = np.where( (train['highest_education'] == 'Post Graduate Qualification'),
                                           'HE Qualification',
                                           train['highest_education']
                                    )


# Do the same for the test set
test['highest_education'] = np.where( (test['highest_education'] == 'No Formal quals'),
                                           'Lower Than A Level',
                                           test['highest_education']
                                    )

test['highest_education'] = np.where( (test['highest_education'] == 'Post Graduate Qualification'),
                                           'HE Qualification',
                                           test['highest_education']
                                    )

Change Age Categories

In [85]:
# Have a closer look at the category

from plotnine import geom_bar, geom_text, theme, element_text
(
    ggplot(train)
    + aes(x='age_band', fill='age_band')
    + geom_bar()
    + geom_text(
     aes(label='stat(prop)*100', group=1),
     stat='count',
     nudge_y=0.125,
     va='bottom',
     format_string='{:.1f}%'
)
    + theme(axis_text_x=element_text(rotation=45, hjust=1))
)
Out[85]:
<Figure Size: (640 x 480)>

Now merge the least frequent categories

In [86]:
# Replace 55+ and 35-55 groups with 35+
train['age_band'] = np.where( (train['age_band'] == '55<='),
                                           '35+',
                                           train['age_band']
                                    )

train['age_band'] = np.where( (train['age_band'] == '35-55'),
                                           '35+',
                                           train['age_band']
                                    )

# Do the same for the test set
test['age_band'] = np.where( (test['age_band'] == '55<='),
                                           '35+',
                                           test['age_band']
                                    )

test['age_band'] = np.where( (test['age_band'] == '35-55'),
                                           '35+',
                                           test['age_band']
                                    )
In [87]:
# See the changes
(
    ggplot(train)
    + aes(x='age_band', fill='age_band')
    + geom_bar()
    + geom_text(
     aes(label='stat(prop)*100', group=1),
     stat='count',
     nudge_y=0.125,
     va='bottom',
     format_string='{:.1f}%'
)
)
Out[87]:
<Figure Size: (640 x 480)>

Bivariate analysis final score vs other variables numerical

In [88]:
sns.pairplot(train)
Out[88]:
<seaborn.axisgrid.PairGrid at 0x7defe3ecf880>

categorical

In [89]:
train = train.assign(fail_final=train['weighted_score'] < 40)

train.head()
Out[89]:
code_module code_presentation date_registration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability total_click weighted_score late_rate fail_rate fail_final
16563 DDD 2014B -36.0 241 F East Anglian Region A Level or Equivalent 50-60% 0-35 1 70 N 0.0 64.9125 0.142857 0.0 False
16006 DDD 2013J -32.0 261 M Scotland A Level or Equivalent 20-30% 0-35 0 120 N 0.0 14.8625 0.000000 0.0 True
12884 CCC 2014J -21.0 269 M South Region HE Qualification 70-80% 35+ 0 60 N 0.0 61.4300 0.875000 0.0 False
27781 FFF 2014J -124.0 269 M West Midlands Region Lower Than A Level 10-20 0-35 0 60 N 0.0 39.5000 0.181818 0.0 True
25125 FFF 2013J -114.0 268 M South East Region Lower Than A Level 10-20 35+ 0 60 N 0.0 9.2500 0.000000 0.0 True

Code_module

In [90]:
(
    ggplot(train)
    + aes(x='code_module', fill='fail_final')
    + geom_bar(position='fill')
    + ggtitle("Count frequency of different modules by pass rate")
)
Out[90]:
<Figure Size: (640 x 480)>
In [91]:
(
    ggplot(train)
    + aes('code_module', 'weighted_score')
    + geom_boxplot(outlier_color='crimson')
    + ggtitle("Distribution of scores per module")
    + coord_flip()
)
Out[91]:
<Figure Size: (640 x 480)>

It seems that some modules have higher fail rates than others. For example, for module CCC the pass rate is just a little over 50%. The boxplot also reveals some outliers.

Code_presentation

In [92]:
(
    ggplot(train)
    + aes(x='code_presentation', fill='fail_final')
    + geom_bar(position='fill')
)
Out[92]:
<Figure Size: (640 x 480)>
In [93]:
(
    ggplot(train)
    + aes('code_presentation', 'weighted_score')
    + geom_boxplot(outlier_color='crimson')
    + ggtitle("Distribution of scores per presentation")
    + coord_flip()
)
Out[93]:
<Figure Size: (640 x 480)>

Gender

In [94]:
(
    ggplot(train)
    + aes(x='gender', fill='fail_final')
    + geom_bar(position='fill')
)
Out[94]:
<Figure Size: (640 x 480)>
In [95]:
(
    ggplot(train)
    + aes('gender', 'weighted_score')
    + geom_boxplot(outlier_color='crimson')
    + ggtitle("Distribution of scores per sex")
    + coord_flip()
)
Out[95]:
<Figure Size: (640 x 480)>

Region

In [96]:
(
    ggplot(train)
    + aes(x='region', fill='fail_final')
    + geom_bar(position='fill')
    + theme(axis_text_x=element_text(rotation=45, hjust=1))
)
Out[96]:
<Figure Size: (640 x 480)>
In [97]:
(
    ggplot(train)
    + aes('region', 'weighted_score')
    + geom_boxplot(outlier_color='crimson')
    + ggtitle("Distribution of scores per region")
    + coord_flip()
)
Out[97]:
<Figure Size: (640 x 480)>

Highest education

In [98]:
(
    ggplot(train)
    + aes(x='highest_education', fill='fail_final')
    + geom_bar(position='fill')
    + theme(axis_text_x=element_text(rotation=45, hjust=1))
)
Out[98]:
<Figure Size: (640 x 480)>

Imd_band

In [99]:
(
    ggplot(train)
    + aes(x='imd_band', fill='fail_final')
    + geom_bar(position='fill')
    + theme(axis_text_x=element_text(rotation=45, hjust=1))
)
Out[99]:
<Figure Size: (640 x 480)>

AGE_band

In [100]:
(
    ggplot(train)
    + aes(x='age_band', fill='fail_final')
    + geom_bar(position='fill')
)
Out[100]:
<Figure Size: (640 x 480)>

Regression

model preparation

In [101]:
# Separate features from target

'''Training set'''
# Drop target and helper columns
X_train = train.drop(columns=['fail_final', 'weighted_score'])
# Make an array with target
Y_train = train['weighted_score'].copy()

'''Test set'''
# Drop target column
X_test = test.drop(columns=['weighted_score'])
# Make an array with target
Y_test = test['weighted_score'].copy()

X_train.head()
Out[101]:
code_module code_presentation date_registration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability total_click late_rate fail_rate
16563 DDD 2014B -36.0 241 F East Anglian Region A Level or Equivalent 50-60% 0-35 1 70 N 0.0 0.142857 0.0
16006 DDD 2013J -32.0 261 M Scotland A Level or Equivalent 20-30% 0-35 0 120 N 0.0 0.000000 0.0
12884 CCC 2014J -21.0 269 M South Region HE Qualification 70-80% 35+ 0 60 N 0.0 0.875000 0.0
27781 FFF 2014J -124.0 269 M West Midlands Region Lower Than A Level 10-20 0-35 0 60 N 0.0 0.181818 0.0
25125 FFF 2013J -114.0 268 M South East Region Lower Than A Level 10-20 35+ 0 60 N 0.0 0.000000 0.0
In [102]:
'''Make a copy for the subsequent last evaluation'''
X_train_eval = X_train.copy()
Y_train_eval = Y_train.copy()
X_test_eval = X_test.copy()
Y_test_eval = Y_test.copy()
In [103]:
from sklearn.preprocessing import OneHotEncoder, OrdinalEncoder, MinMaxScaler, RobustScaler
from sklearn.compose import make_column_transformer

# Set encoding and scaling instructions
column_transform = make_column_transformer(
    (OneHotEncoder(), ['code_module', 'code_presentation', 'gender', 'region', 'age_band', 'disability']),
    (OrdinalEncoder(), ['highest_education', 'imd_band']),
    (RobustScaler(), ['date_registration', 'module_presentation_length',
                       'num_of_prev_attempts', 'studied_credits', 'total_click', 'late_rate',
                       'fail_rate'])
)

# Apply column transformer to features
X_encoded = column_transform.fit_transform(X_train)
In [104]:
# Have a look at what the scaled and encoded data looks like
pd.DataFrame(X_encoded).head()
Out[104]:
0 1 2 3 4 5 6 7 8 9 ... 29 30 31 32 33 34 35 36 37 38
0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 1.0 ... 0.0 0.0 5.0 0.291667 -0.777778 1.0 0.166667 -0.542725 -0.102857 0.0
1 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 ... 0.0 0.0 2.0 0.347222 -0.037037 0.0 1.000000 -0.542725 -0.360000 0.0
2 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 1.0 7.0 0.500000 0.259259 0.0 0.000000 -0.542725 1.215000 0.0
3 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 ... 0.0 2.0 1.0 -0.930556 0.259259 0.0 0.000000 -0.542725 -0.032727 0.0
4 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 1.0 0.0 ... 0.0 2.0 1.0 -0.791667 0.222222 0.0 0.000000 -0.542725 -0.360000 0.0

5 rows × 39 columns

Models

Linear Regression

Linear regression is most likely going to be a bad model as the data breaks several of the assumptions of this model. The assumptions are as follows: Linear relationship between the target and features. The pair plots show this isn't the case. Multivariate normality - all variables must be normal. The histograms of the numerical variables show that their distributions aren't normal. Little to no multicollinearity - all variables must be independent from each other. The correlation matrix shows that this isn't so. No auto-correlation - when the value of y(x+1) is independent from the value of y(x). Homoscedasticity - residuals must be equal along the regression line. Linear regression can show us an example of a bad model, showing how other models can vastly improve the predictions.

In [105]:
from sklearn.linear_model import LinearRegression
from sklearn.pipeline import make_pipeline

# Setting up the pipeline
lm = LinearRegression()

lm_pipeline = make_pipeline(column_transform, lm)

# Fit the training data
lm_pipeline.fit(X_train, Y_train)
# Predict the training data
lm_train_predictions = lm_pipeline.predict(X_train)
In [106]:
# Now let's evaluate the model
import sklearn.metrics as metrics

def regression_eval(X, y, predictions):
    MSE = metrics.mean_squared_error(y, predictions)
    RMSE = np.sqrt(MSE)
    R2 = metrics.r2_score(y, predictions)
    adj_R2 = 1 - ( (1-R2)*(len(y)-1)/(len(y)-X.shape[1]-1) )

    print("-----------------------")
    print('RMSE is {}'.format(RMSE))
    print('Adjusted R2 score is {}\n'.format(adj_R2))

### For training set ###
print("Model performance for training set:")
regression_eval(X_train, Y_train, lm_train_predictions)

Model performance for training set:
-----------------------
RMSE is 27.00893370960526
Adjusted R2 score is 0.2033920643452145

The above evaluation of the model is using our uses a randomised training and test set split, then calculates RMSE and adjusted-R2. Adjusted-R2 shows the model explains 35% of the total variance in the sample. RMSE of 23.8 shows us the error - predictions are off by 23.8 points. This is a large error considering the fail mark is only 40%.

In [107]:
# Perform cross-validation
from sklearn.model_selection import cross_val_score
scores = cross_val_score(lm, column_transform.fit_transform(X_train), Y_train, cv=10, scoring='neg_mean_squared_error')
lm_rmse_scores = np.sqrt(-scores)

def display_scores(scores):
    print('Scores\t:', scores)
    print('Mean\t:', scores.mean())
    print('SD\t:', scores.std())
display_scores(lm_rmse_scores)

Scores	: [27.46407449 26.5050957  26.48338021 27.72973593 27.24882204 26.8113389
 26.97115476 26.8154955  27.1045002  27.40899831]
Mean	: 27.05425960624247
SD	: 0.39253588018554825

Cross-validation uses stratified k-fold cross-validation which is different from validation with randomised values. The training set is split into a smaller training set and an even smaller validation set. Each of these sets are then used for training and validation sequentially. Cross-validation also shows us the model performs abysmally, just as expected.

Decision Tree

This model uses a very different approach - building a what is essentially a flow chart based on probabilities and likelihoods. It is a simple algorithm with many models using it as a base algorithm (e.g. Random Forest).

In [108]:
from sklearn.tree import DecisionTreeRegressor

Dtree = DecisionTreeRegressor(min_samples_leaf=15, min_samples_split=10, max_features=13)

Dtree_pipeline = make_pipeline(column_transform, Dtree)

Dtree_pipeline.fit(X_train, Y_train)

train_Dtree_predictions = Dtree_pipeline.predict(X_train)

### Evaluating the model ###
print("Model performance for training set:")
regression_eval(X_train, Y_train, train_Dtree_predictions)

Model performance for training set:
-----------------------
RMSE is 22.16203045126824
Adjusted R2 score is 0.4636493081995984
In [109]:
scores = cross_val_score(Dtree, column_transform.fit_transform(X_train), Y_train, cv=5, scoring='neg_mean_squared_error')

Dtree_rmse_scores = np.sqrt(-scores)
display_scores(Dtree_rmse_scores)

Scores	: [25.42317994 25.78908881 26.07331791 25.94282216 25.39452497]
Mean	: 25.72458676063109
SD	: 0.27319863737156963

K Nearest Neighbours Regression

KNN-Regression examines the point close to the target point and then makes a prediction on which class these data points belong to.

In [110]:
from sklearn.neighbors import KNeighborsRegressor

KNReg = KNeighborsRegressor(n_neighbors=2)

KNReg_pipeline = make_pipeline(column_transform, KNReg)

KNReg_pipeline.fit(X_train, Y_train)

train_KNReg_predictions = KNReg_pipeline.predict(X_train)

### Evaluating the model ###
print("Model performance for training set:")
regression_eval(X_train, Y_train, train_KNReg_predictions)

Model performance for training set:
-----------------------
RMSE is 19.247816461915214
Adjusted R2 score is 0.5954309197456498
In [111]:
scores = cross_val_score(KNReg, column_transform.fit_transform(X_train), Y_train, cv=5, scoring='neg_mean_squared_error')

KNReg_rmse_scores = np.sqrt(-scores)
display_scores(KNReg_rmse_scores)

Scores	: [33.29889186 33.34006999 33.47508103 33.50954947 33.61278295]
Mean	: 33.44727506036353
SD	: 0.1145029560403488

Random Forest

In [112]:
from sklearn.ensemble import RandomForestRegressor

RForest = RandomForestRegressor(min_samples_leaf=15, min_samples_split=10,
                                max_features=13, n_estimators=20)

RForest_pipeline = make_pipeline(column_transform, RForest)

RForest_pipeline.fit(X_train, Y_train)

train_RForest_predictions = RForest_pipeline.predict(X_train)

### Evaluating the model ###
print("Model performance for training set:")
regression_eval(X_train, Y_train, train_RForest_predictions)

Model performance for training set:
-----------------------
RMSE is 21.84270847046474
Adjusted R2 score is 0.4789939944829905
In [113]:
scores = cross_val_score(RForest, column_transform.fit_transform(X_train), Y_train, cv=5, scoring='neg_mean_squared_error')

RForest_rmse_scores = np.sqrt(-scores)
display_scores(RForest_rmse_scores)

Scores	: [23.60455172 23.87533037 23.82154874 23.87346647 23.89264068]
Mean	: 23.81350759367239
SD	: 0.107150721045399

support vector regression

Linear models are doing badly, as expected. We will next use non-linear models and see if our predictions improve. Support Vector Regression seeks not to minimise the squared error as in the linear regression, but to minimise coefficients.

In [114]:
from sklearn.svm import SVR

SVR = SVR(kernel='rbf')

SVR_pipeline = make_pipeline(column_transform, SVR)

SVR_pipeline.fit(X_train, Y_train)

train_SVR_predictions = SVR_pipeline.predict(X_train)

### Evaluating the model ###
print("Model performance for training set:")
regression_eval(X_train, Y_train, train_SVR_predictions)

Model performance for training set:
-----------------------
RMSE is 27.58378874549837
Adjusted R2 score is 0.16912137475949185
In [115]:
scores = cross_val_score(SVR, column_transform.fit_transform(X_train), Y_train, cv=4, scoring='neg_mean_squared_error')

SVR_rmse_scores = np.sqrt(-scores)
display_scores(SVR_rmse_scores)

Scores	: [27.60848003 27.97666461 27.56569268 28.28110677]
Mean	: 27.85798602203321
SD	: 0.29189228133113526

The model is an improvement on the linear models. Adjusted R2 is a bit higher (39%) and the standard error for RMSE scores in the cross-validated sets is lower (from SD = 0.39 for Random Forest to SD = 0.10). It is still not a great predictor for the dataset.

Clasification model

model preparation

In [119]:
train = train_class.copy()
test = test_class.copy()

train.head()
Out[119]:
code_module code_presentation date_registration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability final_result total_click
16563 DDD 2014B -36.0 241 F East Anglian Region A Level or Equivalent 50-60% 0-35 1 70 N Pass 0.0
16006 DDD 2013J -32.0 261 M Scotland A Level or Equivalent 20-30% 0-35 0 120 N Withdrawn 0.0
12884 CCC 2014J -21.0 269 M South Region HE Qualification 70-80% 55<= 0 60 N Pass 0.0
27781 FFF 2014J -124.0 269 M West Midlands Region Lower Than A Level 10-20 0-35 0 60 N Fail 0.0
25125 FFF 2013J -114.0 268 M South East Region Lower Than A Level 10-20 35-55 0 60 N Fail 0.0

Last Cleaning

In [120]:
# Distinction as a Pass
train['final_result'] = np.where( (train['final_result'] == 'Distinction'),
                                           'Pass',
                                           train['final_result']
                                    )
# Withdrawn as a Fail (to make the target binary)
train['final_result'] = np.where( (train['final_result'] == 'Withdrawn'),
                                           'Fail',
                                           train['final_result']
                                    )
# Same for test set
test['final_result'] = np.where( (test['final_result'] == 'Distinction'),
                                           'Pass',
                                           test['final_result']
                                    )
test['final_result'] = np.where( (test['final_result'] == 'Withdrawn'),
                                           'Fail',
                                           test['final_result']
                                    )
In [121]:
# Rename 'no formal quals' into 'lower than a level'
train['highest_education'] = np.where( (train['highest_education'] == 'No Formal quals'),
                                           'Lower Than A Level',
                                           train['highest_education']
                                    )

# Rename post-grads
train['highest_education'] = np.where( (train['highest_education'] == 'Post Graduate Qualification'),
                                           'HE Qualification',
                                           train['highest_education']
                                    )


# Do the same for the test set
test['highest_education'] = np.where( (test['highest_education'] == 'No Formal quals'),
                                           'Lower Than A Level',
                                           test['highest_education']
                                    )

test['highest_education'] = np.where( (test['highest_education'] == 'Post Graduate Qualification'),
                                           'HE Qualification',
                                           test['highest_education']
                                    )
### Age bands ###
train['age_band'] = np.where( (train['age_band'] == '55<='),
                                           '35-55',
                                           train['age_band']
                                    )

train['age_band'] = np.where( (train['age_band'] == '35-55'),
                                           '35+',
                                           train['age_band']
                                    )

# Do the same for the test set
test['age_band'] = np.where( (test['age_band'] == '55<='),
                                           '35-55',
                                           test['age_band']
                                    )

test['age_band'] = np.where( (test['age_band'] == '35-55'),
                                           '35+',
                                           test['age_band']
                                    )
In [122]:
# Separate features from target

'''Training set'''
# Drop target column
X_train = train.drop(columns=['final_result'])
# Make an array with target
Y_train = train['final_result'].copy()

'''Test set'''
# Drop target column
X_test = test.drop(columns=['final_result'])
# Make an array with target
Y_test = test['final_result'].copy()

X_train.head()
Out[122]:
code_module code_presentation date_registration module_presentation_length gender region highest_education imd_band age_band num_of_prev_attempts studied_credits disability total_click
16563 DDD 2014B -36.0 241 F East Anglian Region A Level or Equivalent 50-60% 0-35 1 70 N 0.0
16006 DDD 2013J -32.0 261 M Scotland A Level or Equivalent 20-30% 0-35 0 120 N 0.0
12884 CCC 2014J -21.0 269 M South Region HE Qualification 70-80% 35+ 0 60 N 0.0
27781 FFF 2014J -124.0 269 M West Midlands Region Lower Than A Level 10-20 0-35 0 60 N 0.0
25125 FFF 2013J -114.0 268 M South East Region Lower Than A Level 10-20 35+ 0 60 N 0.0
In [123]:
X_train.shape
Out[123]:
(26074, 13)

Encoding for trees

In [124]:
from sklearn.preprocessing import OneHotEncoder, OrdinalEncoder, StandardScaler, MinMaxScaler, RobustScaler
from sklearn.compose import make_column_transformer

# Set encoding and scaling instructions
column_transform = make_column_transformer(
    (OneHotEncoder(), ['code_module', 'code_presentation', 'gender', 'region', 'disability']),
    (OrdinalEncoder(), ['highest_education', 'imd_band', 'age_band']),
    remainder='passthrough'
)

# Apply column transformer to features
X_encoded = column_transform.fit_transform(X_train)
In [125]:
pd.DataFrame(X_encoded).head()
Out[125]:
0 1 2 3 4 5 6 7 8 9 ... 26 27 28 29 30 31 32 33 34 35
0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 1.0 ... 1.0 0.0 0.0 5.0 0.0 -36.0 241.0 1.0 70.0 0.0
1 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 ... 1.0 0.0 0.0 2.0 0.0 -32.0 261.0 0.0 120.0 0.0
2 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 1.0 0.0 1.0 7.0 1.0 -21.0 269.0 0.0 60.0 0.0
3 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 ... 1.0 0.0 2.0 1.0 0.0 -124.0 269.0 0.0 60.0 0.0
4 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 1.0 0.0 ... 1.0 0.0 2.0 1.0 1.0 -114.0 268.0 0.0 60.0 0.0

5 rows × 36 columns

Models

Decision tree

In [129]:
from sklearn.tree import DecisionTreeClassifier

Dtree = DecisionTreeClassifier(min_samples_leaf=15, min_samples_split=10, max_features=8)

Dtree_pipeline = make_pipeline(column_transform, Dtree)

# Cross-validate
def display_accuracy_scores(pipeline, X, Y):
    scores = cross_val_score(pipeline, X, Y, cv=5, scoring='accuracy')
    print('Scores\t:', scores)
    print('Mean\t:', scores.mean())
    print('SD\t:', scores.std())

### Cross-validate ###
# Train set
print('Evaluation of the training set')
display_accuracy_scores(Dtree_pipeline, X_train, Y_train)

Evaluation of the training set
Scores	: [0.61093001 0.61457335 0.62013423 0.60460211 0.60663598]
Mean	: 0.611375134649283
SD	: 0.005576412308870381

Random forest

In [130]:
from sklearn.ensemble import RandomForestClassifier

RForest = RandomForestClassifier(min_samples_leaf=15, min_samples_split=10,
                                max_features=8, n_estimators=20)

RForest_pipeline = make_pipeline(column_transform, RForest)

### Cross-validate ###
# Train set
print('Evaluation of the training set')
display_accuracy_scores(RForest_pipeline, X_train, Y_train)

Evaluation of the training set
Scores	: [0.64793864 0.6506232  0.65292426 0.64391179 0.64902186]
Mean	: 0.6488839509823283
SD	: 0.002999468589052481

Random forest classifier model performed the best. Althrough its accuracy scores (0.64, SD = 0.002) .

In [131]:
# Test set evaluation for SVC
print('Evaluation of the test set')

# Fit the training data
SVClass_pipeline.fit(X_train, Y_train)
# Transform the test data
X_test_prepared = column_transform.transform(X_test)
# Predict the test data
SVClass_predictions_test = SVClass.predict(X_test_prepared)

print('Accuracy score:', metrics.accuracy_score(Y_test, SVClass_predictions_test))

Evaluation of the test set
Accuracy score: 0.6458045712532597

To quickly summarise, the best model for the regression task was support vector (RMSE = 18.4, SD = 0.2 with 4-fold cross-validation) which gave us RMSE = 18.1 and adjusted R2 = 0.63 when evaluated on the test set). This is without any fine-tuning of the hyperparameters. The best model for the classification task was a Random forest (0.78% accuracy score on the test set). Again, this is without any hyperparameter tuning. linkcode Next steps to take would be to find which features are most important and which can be dropped. Hyperparameter tuning can be used to find the best set of parameters for the models. Various dimensionality reduction tools can be used to improve the performance of the models. Another point to make is that accuracy score isn't the best way to evaluate classification models, especially when the target is imbalanced. We have an imbalanced target for this classification problem, so dealing with this imbalance and using a different evaluation metric would be advantageous.